Thomas Capelle

Results 169 comments of Thomas Capelle

@andrewtruong as I added a method to dataset, is it normal that the hashes changes? (and thus the failing tests)

> Yes. Since you added a new `op`, the digest is expected to change How do I fix the tests then?

If I randomly print text on the model it doesn't disturb the progress:

Is this still necessary @andrewtruong ?

![Screenshot 2025-03-17 at 16 52 39](https://github.com/user-attachments/assets/796a204a-f341-4ceb-8166-b4575e17aff6) we do get weird behavior here

I still prefer this weird behavior than what we have now.

Can we merge this @andrewtruong ?

You could add a section about the flow: # GRPO Repeated Sampling Flow Implementation This document explains how this GRPO implementation generates multiple different completions for each prompt when `num_generations...